From 6b7cc8a9ecb37e712679419008fb661e832abf49 Mon Sep 17 00:00:00 2001 From: Jason Ekstrand Date: Mon, 7 Nov 2016 12:32:28 -0800 Subject: [PATCH] anv: Document cmd_buffer_alloc_binding_table Some of the details of this function are very confusing and have a long history. We should document that history and this seems like the best place to do it. Signed-off-by: Jason Ekstrand --- src/intel/vulkan/anv_batch_chain.c | 71 ++++++++++++++++++++++++++++++ 1 file changed, 71 insertions(+) diff --git a/src/intel/vulkan/anv_batch_chain.c b/src/intel/vulkan/anv_batch_chain.c index dfa9abfabaa..1e348cfb004 100644 --- a/src/intel/vulkan/anv_batch_chain.c +++ b/src/intel/vulkan/anv_batch_chain.c @@ -522,6 +522,77 @@ anv_cmd_buffer_grow_batch(struct anv_batch *batch, void *_data) return VK_SUCCESS; } +/** Allocate a binding table + * + * This function allocates a binding table. This is a bit more complicated + * than one would think due to a combination of Vulkan driver design and some + * unfortunate hardware restrictions. + * + * The 3DSTATE_BINDING_TABLE_POINTERS_* packets only have a 16-bit field for + * the binding table pointer which means that all binding tables need to live + * in the bottom 64k of surface state base address. The way the GL driver has + * classically dealt with this restriction is to emit all surface states + * on-the-fly into the batch and have a batch buffer smaller than 64k. This + * isn't really an option in Vulkan for a couple of reasons: + * + * 1) In Vulkan, we have growing (or chaining) batches so surface states have + * to live in their own buffer and we have to be able to re-emit + * STATE_BASE_ADDRESS as needed which requires a full pipeline stall. In + * order to avoid emitting STATE_BASE_ADDRESS any more often than needed + * (it's not that hard to hit 64k of just binding tables), we allocate + * surface state objects up-front when VkImageView is created. In order + * for this to work, surface state objects need to be allocated from a + * global buffer. + * + * 2) We tried to design the surface state system in such a way that it's + * already ready for bindless texturing. The way bindless texturing works + * on our hardware is that you have a big pool of surface state objects + * (with its own state base address) and the bindless handles are simply + * offsets into that pool. With the architecture we chose, we already + * have that pool and it's exactly the same pool that we use for regular + * surface states so we should already be ready for bindless. + * + * 3) For render targets, we need to be able to fill out the surface states + * later in vkBeginRenderPass so that we can assign clear colors + * correctly. One way to do this would be to just create the surface + * state data and then repeatedly copy it into the surface state BO every + * time we have to re-emit STATE_BASE_ADDRESS. While this works, it's + * rather annoying and just being able to allocate them up-front and + * re-use them for the entire render pass. + * + * While none of these are technically blockers for emitting state on the fly + * like we do in GL, the ability to have a single surface state pool is + * simplifies things greatly. Unfortunately, it comes at a cost... + * + * Because of the 64k limitation of 3DSTATE_BINDING_TABLE_POINTERS_*, we can't + * place the binding tables just anywhere in surface state base address. + * Because 64k isn't a whole lot of space, we can't simply restrict the + * surface state buffer to 64k, we have to be more clever. The solution we've + * chosen is to have a block pool with a maximum size of 2G that starts at + * zero and grows in both directions. All surface states are allocated from + * the top of the pool (positive offsets) and we allocate blocks (< 64k) of + * binding tables from the bottom of the pool (negative offsets). Every time + * we allocate a new binding table block, we set surface state base address to + * point to the bottom of the binding table block. This way all of the + * binding tables in the block are in the bottom 64k of surface state base + * address. When we fill out the binding table, we add the distance between + * the bottom of our binding table block and zero of the block pool to the + * surface state offsets so that they are correct relative to out new surface + * state base address at the bottom of the binding table block. + * + * \see adjust_relocations_from_block_pool() + * \see adjust_relocations_too_block_pool() + * + * \param[in] entries The number of surface state entries the binding + * table should be able to hold. + * + * \param[out] state_offset The offset surface surface state base address + * where the surface states live. This must be + * added to the surface state offset when it is + * written into the binding table entry. + * + * \return An anv_state representing the binding table + */ struct anv_state anv_cmd_buffer_alloc_binding_table(struct anv_cmd_buffer *cmd_buffer, uint32_t entries, uint32_t *state_offset) -- 2.30.2